An investigation into the application of incremental capacity analysis in harsh measurement conditions: focusing on the VC model-based ICA technique and the influence of training set selection on the SOH model

Applying incremental capacity analysis (ICA) under harsh measurement conditions warrants further investigation. Although partial- and variable-charging ICA studies have improved practical applicability, the coupled influence of variable CC-charging segments and degraded measurement quality on ICA-based SOH estimation remains insufficiently explored. A notable research gap, therefore, remains in determining how to construct training sets for accurate and robust SOH estimation, particularly under elevated noise levels and low sampling rates. This work utilizes the established Revised Lorentzian-Voltage Capacity (RL-VC) model as a parameter-level ICA representation and evaluates its robustness to noisy, sparsely sampled measurements. A comparison between RL-VC-based and filter-based ICA across various noise levels and sampling rates is presented. To facilitate this comparison, variable CC-charging segments with different starting SOCs and charging rates are used to generate the training and test sets necessary for developing the SOH model. For the first time, this study integrates training set selection to propose a robustness-optimized method for SOH modeling. Three distinct model structures (feed-forward neural network, FFNN, Gaussian process regression, GPR, and support vector regression, SVR) are utilized for the modeling process to substantiate the conclusions drawn. Under variable charging with high-quality measurements, the RL-VC-based SOH estimators achieve RMSEs of 1.20% (FFNN) and 1.10% (GPR), outperforming a filter-based ICA baseline (1.53% RMSE) under the same conditions. The findings indicate that the RL-VC approach demonstrates considerably superior robustness to the filter-based scheme under harsh measurement conditions. Notably, even with the introduction of Gaussian noise with a standard deviation of up to 10% or sampling intervals extending to 60 s, the RL-VC approach consistently produces precise IC curves. The analysis of variance (ANOVA) results demonstrates that the selection of the training set substantially impacts the robustness of the SOH model under harsh conditions. Across the harsh-condition scenario matrix, the lowest average RMSE is 1.47% when the training data are augmented with 5% Gaussian noise (GPR). Within the defined Gaussian-noise and fixed-sampling scenario matrix, combining a 30 s sampling interval with 5% Gaussian noise in the training set provides a robust empirical trade-off, yielding an average RMSE of 1.49% (GPR).